Gradient descent is hugely controversial in neuroscience. See the ultimate megathread: https://treeverse.app/view/1hX6Xm0w
Prikaži ovu nit
-
Frankly, I am a bit puzzled by most counter-arguments. Our best physical theories are mathematical *descriptions* of reality. If you take Newton’s inverse-square law for example this describes experimental phenomena but no astronomer thinks that planets do calculations.
There are several lines of reasoning to justify the gradient formalism. First, the argument of practicality. We now have very good automatic differentiation software that can revolutionise the theoretical sciences. Why don’t theoretical neuroscientists make good use of this?
-
-
Second, the argument of usefulness. Arguably most learning happens during the first couple years of brain development. During this period the brain is reorganised on multiple spatial and temporal scales. This physical process can be described using automatic differentiation.
-
All we want as scientists are compact descriptions of reality that may be used to make predictions. These models are observer-dependent so there appears to be confusion about *physiological* mechanisms. Our understanding of physiology is constrained by our instruments.
- Još 1 odgovor
Novi razgovor -
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.